How to use unsupervised machine learning to help identify corneal severity

Abstract

To provide a method for assisting corneal severity identification using unsupervised machine learning or a method for identifying normal, disorder-suspect and disorder eyes by applying machine learning to a number of corneal tomography data.SOLUTION: A method for assisting corneal severity identification comprises: obtaining a corneal configuration data set of a cornea examined by a tomography such as an optical coherence tomography; visualizing the corneal configuration data set of the cornea examined along with a number of pre-existing corneal configuration data sets of disorder corneas, disorder-suspect corneas and normal corneas obtained by the tomography using t-distributed Stochastic Neighbor Embedding in a two or three dimensional map; and judging corneal severity from the map.SELECTED DRAWING: Figure 3

Description

[Technical field]

[0001] The present invention relates to a method for using unsupervised machine learning to aid in the identification of corneal severity. [Background technology]

[0002] Corneal diseases such as keratoconus, bullous keratopathy, walleye, keratoleukoma, ulcer, herpes, chemical burn, thermal injury, and degenerative corneal degeneration can be cured by corneal transplantation. Determining the need for a corneal transplant is difficult. Keratoconus is a non-inflammatory ectatic corneal disease characterized by gradual thinning of the cornea, resulting in corneal protrusion and impaired visual acuity (Non-Patent Document 1). Moderate to advanced keratoconus cases are easy to diagnose due to the presence of typical symptoms by keratoscopic examination and biomicroscopy. However, detection of latent keratoconus is difficult because early signs of keratoconus can be unclear and require a more comprehensive analysis of corneal characteristics, including topography, elevation, thickness, and biomechanical properties (Non-Patent Documents 2 and 3). Many methods have been proposed to identify keratoconic eyes using corneal topography information. However, most of the methods rely on subjective analysis of topographic maps, which may bias the observer (Non-Patent Document 4).

[0003] Among the objective approaches to the identification of keratoconus, machine learning analysis has attracted much attention. Smolek and Klyce (Non-Patent Document 5) proposed a neural network for keratoconus screening based on corneal topography index. Chastang et al. (Non-Patent Document 6) introduced a binary decision tree method based on corneal topography index to identify clinically evident keratoconus from normal corneas. A few years later, a similar approach was used to identify keratoconus from normal corneas, using a corneal surface modeled by a 7th order Zernike polynomial (Non-Patent Document 7). All these methods used only the anterior topography characteristics of the cornea. However, with the advancement of technology, posterior corneal curvature and pachymetry data are acquired and used to evaluate corneal characteristics (Non-Patent Document 8). Pinero et al. provided documentation of the correlation between corneal volume, pachymetry, and anterior and posterior corneal shape in latent and clinical keratoconus (Non-Patent Document 9). Perez et al. showed that the use of corneal instruments including video corneal topography measuring devices such as Orbscan and Pentacam in conjunction with the index can result in early detection of keratoconus, but increases false positive detection (Non-Patent Document 10).

[0004] Current methods for automatic detection of keratoconus are primarily supervised machine learning, in the sense that they require labeling and diagnosis as input to the next machine learning. The present invention proposes a method that does not bias the clinician or the patient. The present method can result in a better identification of incomplete keratoconus, which in some cases may be difficult to identify clinically. Furthermore, the present method provides an unbiased way to determine progression and the need for other treatments, such as correlation (Non-Patent Document 11). From a big data perspective, the proposed method is objective and does not require the eye to be labelled beforehand. The present results suggest that unsupervised machine learning can be applied to corneal topography, elevation and pachymetry parameters to generate highly specific and sensitive models.

[0005] Furthermore, Patent Document 1 discloses a method for diagnosing keratoconus in a patient's eye, the method comprising the steps of: providing an electronic model of the cornea, the model including a plurality of elements, each element in the model being defined by a plurality of parameters, each parameter representing the quality of the corneal tissue at a corresponding location of the element in the model; mapping a topography "Ta" of the anterior corneal surface; fitting the topography "Ta" to the model to obtain a parameter set of the plurality of elements; and evaluating the parameter set and diagnosing whether the cornea is keratoconus.

[0006] Patent document 2 discloses a method for measuring the cornea of ​​an eye, the method comprising: using an ophthalmic analysis system to measure a corneal topography undergoing intraocular pressure changes during a measurement period; obtaining several image data sets of a surface area of ​​the cornea during the measurement period, which cause repeated changes in the corneal topography during the measurement period due to the intraocular pressure changes during the measurement period; and determining from the several image data the repeated changes in the corneal topography, in each case the changes being measured at a point (P) of the corneal surface area measured during the topography measurement.

[0007] Patent document 3 discloses a method for analyzing corneal topography, the method including the steps of obtaining corneal curvature data; determining a plurality of indices characterizing the corneal topography based on the obtained corneal curvature data; and using the determined indices and a neural network to determine the corneal topography from features specific to a predetermined classification of the corneal topography and determine at least one of normal cornea, myopic refractive surgery, hyperopic refractive surgery, corneal astigmatism, penetrating keratoplasty, keratoconus, suspected keratoconus, pellucid marginal degeneration or other classification of corneal topography. [Prior art documents] [Patent documents]

[0008] [Patent Document 1] Japanese Patent Application Laid-Open No. 2010-57897 (US2010 / 006854A1) [Patent Document 2] Japanese Patent Application Laid-Open No. 2013-255791 (US2013 / 0329187A1) [Patent Document 3] Japanese Patent Application Laid-Open No. 2005-288176 (US2005 / 225724A1) [Non-Patent Document]

[0009] [Non-Patent Document 1] Rabinowitz YS. Keratoconus. Surv Ophthalmol 1998;42:297~319 [Non-Patent Document 2] de Sanctis U, Loiacono C, Richiardi L, Turco D, Mutani B, Grignolo FM. Sensitivity and specificity of posterior corneal elevation measured by Pentacam in discriminating keratoconus / subclinical keratoconus. Ophthalmology 2008;115:1534~1539 [Non-Patent Document 3] Gordon-Shaag A, Millodot M, Ifrah R, Shneor E. Aberrations and topography in normal, keratoconus-suspect, and keratoconic eyes. Optom Vis Sci 2012;89:411~418 [Non-Patent Document 4] Maeda N, Klyce SD, Smolek MK, Thompson HW. Automated keratoconus screening with corneal topography analysis. Invest Ophthalmol Vis Sci 1994;35:2749~2757 [Non-Patent Document 5] Smolek MK, Klyce SD. Current keratoconus detection methods compared with a neural network approach. Invest Ophthalmol Vis Sci 1997;38:2290~2299

Non-Patent Document 6

Non-Patent Document 7

Non-Patent Document 8

Non-Patent Document 9

Outdoor Tools 10

Outdoor Content11

Outdoor Tools 12

Outdoor Tools13

Non-Patent Document 18

Non-Patent Document 19

Non-Patent Document 20

Direct Environment21

Outdoor Tools22

Optional Trademark23

Optional Trademark24

Direct Entries 25

Direct Entries 26

Direct Accounts 27

Direct Account 28

Direct Environment 29

Draw 30 pages

[0010] It is an object of the present invention to provide a method to use unsupervised machine learning to assist in identifying corneal severity or to identify normal, suspect and diseased eyes by applying machine learning to large numbers of corneal tomography data. [Means for solving the problem]

[0011] The present invention is directed to a method for aiding in the identification of corneal severity.

[0012] In one embodiment of the present invention, the method includes obtaining a corneal configuration dataset of a cornea examined by tomography; visualizing the corneal configuration dataset of the examined cornea in a two-dimensional or three-dimensional map using t-distributed Stochastic Neighbor Embedding together with several existing corneal configuration datasets of diseased, suspected and normal corneas obtained by tomography; and determining corneal severity from the map. In another embodiment of the invention, the corneal severity is at least one selected from the group consisting of keratoconus severity, bullous keratopathy severity, walleye severity, keratoleukoma severity, keratohelcosis severity, herpes corneae severity, corneal chemical burn severity, corneal thermal burn severity, and degeneratio corneae severity.

[0013] In another embodiment of the present invention, the corneal configuration data set includes at least one of a 2D analysis of the ACA observation plane, a 2D analysis of the CCT / ACD observation plane, STAR360° (Scleral Spur Tracking for Angle Analysis And Registration 360°), lens morphology analysis in 2D results, lens morphology analysis in 3D results, and corneal morphology analysis and reference points.

[0014] In another embodiment of the invention, the 2D analysis of the ACA viewing plane includes at least one of AOD500, AOD750, ARA500, ARA750, TISA500, TISA750, TIA500 and TIA750 (FIG. 6).

[0015] AOD500 (Angle Opening Distance 500) is the distance (mm) between AOD500-T and AOD500-IF, AOD500-T is a point on the trabecular meshwork (on the inner surface of the cornea) 500 μm away from the scleral spur (SS), AOD500-IF is the point where a line that intersects perpendicularly with a line connecting the SS and ARA-T and passes through AOD500-T intersects with the anterior surface of the iris, and ARA-T is a point on the trabecular meshwork (on the inner surface of the cornea) 750 μm away from the SS.

[0016] AOD750 (Angle Opening Distance 750) is the distance (mm) between ARA-T and ARA-IF, ARA-T is a point on the trabecular meshwork (on the inner surface of the cornea) 750 μm away from SS, and ARA-IF is the point where a line that is perpendicular to the line connecting SS and ARA-T and passes through ARA-T intersects with the anterior surface of the iris.

[0017] ARA500 (Angle Recess Area 500) is the area (mm 2 ) of the bottom of the angle enclosed by the line connecting AOD500-T and AOD500-IF. 2 ).

[0018] ARA750 (Angle Recess Area 750) is the area (mm 2 ) of the angle recess area enclosed by the line connecting ARA-T and ARA-IF. 2 ).

[0019] TISA500 (Trabecular Iris Space Area 500) is the area (mm 2 ) of the corner base region enclosed by the line connecting SS and SS-IF and the line connecting AOD500-T and AOD500-IF. 2), and SS-IF is the point where a line passing through SS and perpendicular to a line connecting SS and ARA-T intersects with the anterior surface of the iris.

[0020] TISA750 (Trabecular Iris Space Area 750) is the area (mm ) of the iris base enclosed by the line connecting SS and SS-IF and the line connecting ARA-T and ARA-IF. 2 ).

[0021] TIA500 (Trabecular-Iris Angle 500) is the angle (°) between a line connecting AR and AOD500-T and a line connecting AR and AOD500-IF, where AR is the base of the iris angle.

[0022] TIA750 (Trabecular-Iris Angle 750) is the angle (°) between the line connecting AR and ARA-T and the line connecting AR and ARA-IF.

[0023] In another embodiment of the present invention, the 2D analysis of the CCT / ACD view plane includes at least one of CCT, ACD Endo., LV, ACW, CCT, ACD[Epi.], ACD[Endo.], Vault, CLR, and ATA.

[0024] CCT (Central Corneal Thickness) is the corneal thickness (μm) at the point where the perpendicular bisector of the ACW connecting the scleral scarps (SSs) intersects with the cornea.

[0025] ACD Endo. (Anterior Chamber Depth Endothelium) is the anterior chamber depth (mm) from the posterior surface of the cornea to the anterior surface of the lens.

[0026] LV (Lens Vault) is the distance (mm) between the anterior surface of the lens on the perpendicular bisector of the ACW line connecting the scleral promontory (SS) and the perpendicular bisecting point of the ACW.

[0027] ACW (Anterior Chamber Width) is the distance (mm) between the scleral promontory (SS).

[0028] CCT (Central Corneal Thickness) is the corneal thickness (μm) at the point where the perpendicular bisector of the line segment connecting the angle bases (AR) intersects with the cornea.

[0029] ACD[Epi.] (Anterior Chamber Depth [Epithelium]) is the anterior chamber depth (mm) from the anterior surface of the cornea to the anterior surface of the lens.

[0030] ACD[Endo.] (Anterior Chamber Depth [Endothelium]) is the anterior chamber depth (mm) from the posterior surface of the cornea to the anterior surface of the lens.

[0031] Vault (Vault) is the distance (μm) from the posterior surface of the phakic lens (IOL) to the anterior surface of the lens.

[0032] CLR (Crystalline Lens Rise) is the distance (μm) between the anterior surface of the lens on the perpendicular bisector of the line segment (ATA) connecting the iris bases (AR) and the perpendicular bisector of the ATA.

[0033] ATA (Angle to Angle) is the distance (mm) between the angulation angles (AR).

[0034] In another embodiment of the present invention, STAR360° (Scleral Spur Tracking For Angle Analysis And Registration 360°) includes at least one of AOD250, AOD500, AOD750, ARA250, ARA500, ARA750, TISA250, TIA500, TIA750, CCT, ACD Endo., LV, ACW, AC.Area, IT750, IT2000, I-Curv. and ITC.

[0035] AOD250 (Angle Opening Distance 250) is the distance (mm) between AOD250-T and AOD250-IF, AOD250-T is a point on the trabecular meshwork (on the inner surface of the cornea) 250 μm away from the scleral spur (SS), AOD250-IF is the point where a line that is perpendicular to the line connecting the SS and ARA-T and passing through AOD250-T intersects with the anterior surface of the iris, and ARA-T is a point on the trabecular meshwork (on the inner surface of the cornea) 750 μm away from the scleral spur (SS).

[0036] AOD500 (Angle Opening Distance 500) is the distance (mm) between AOD500-T and AOD500-IF, AOD500-T is a point on the trabecula (on the inner surface of the cornea) 500 μm away from the SS, AOD500-IF is the point where a line that is perpendicular to the line connecting SS and ARA-T and passing through AOD500-T intersects with the anterior surface of the iris, and ARA-T is a point on the trabecula (on the inner surface of the cornea) 750 μm away from the SS.

[0037] AOD750 (Angle Opening Distance 700) is the distance (mm) between ARA-T and ARA-IF, ARA-T is a point on the trabecular meshwork (on the inner surface of the cornea) 750 μm away from SS, and ARA-IF is the point where a line that is perpendicular to the line connecting SS and ARA-T and passing through ARA-T intersects with the anterior surface of the iris.

[0038] ARA250 (Angle Recess Area 250) is the area (mm ) of the corner recess area enclosed by the line connecting AOD250-T and AOD250-IF. 2 ).

[0039] ARA500 (Angle Recess Area 500) is the area (mm ) of the corner recess area enclosed by the line connecting AOD500-T and AOD500-IF. 2 ).

[0040] ARA750 (Angle Recess Area 750) is the area (mm ) of the corner recess area enclosed by the line connecting ARA-T and ARA-IF. 2 ).

[0041] TISA250 (Trabecular Iris Space Area 250) is the area (mm2) of the corner base region enclosed by the line connecting SS and SS-IF and the line connecting AOD250-T and AOD250-IF. 2 ), and SS-IF is the point where a line that is perpendicular to a line connecting SS and ARA-T and passes through SS intersects with the anterior surface of the iris.

[0042] TIA500 (Trabecular-Iris Angle 500) is the angle (°) between a line connecting AR and AOD500-T and a line connecting AR and AOD500-IF, where AR is the base of the iris.

[0043] TIA750 (Trabecular-Iris Angle 750) is the angle (°) between the line connecting AR and ARA-T and the line connecting AR and ARA-IF.

[0044] CCT (Central Corneal Thickness) is the corneal thickness (μm) at the point where the perpendicular bisector of the ACW line connecting the scleral promontory (SS) intersects with the cornea.

[0045] ACD Endo. (Anterior Chamber Depth Endothelium) is the anterior chamber depth (mm) from the posterior surface of the cornea to the anterior surface of the lens.

[0046] LV (Lens Vault) is the distance (mm) between the anterior surface of the lens on the perpendicular bisector of the ACW line connecting the scleral promontory (SS) and the perpendicular bisecting point of the ACW.

[0047] ACW (Anterior Chamber Width) is the distance (mm) between the scleral promontory (SS).

[0048] AC.Area (Anterior Chamber Area) is the area of ​​the anterior chamber (mm 2 ).

[0049] IT750 (Iris Thickness 750) is the thickness (mm) of the iris at a position 750 μm away from the SS.

[0050] IT2000 (Iris Thickness 2000) is the thickness (mm) of the iris at a position 2000 μm away from the SS.

[0051] I-Area (Iris Area) is the area of ​​the iris (mm 2 ).

[0052] I-Curv. (Iris Curvatura) is the maximum distance (mm) from the line connecting the iris root (IR) and the lens contact end point (IRT) to the iris pigment epithelium (back surface).

[0053] ITC (Irido-Trabecular Contact) is the contact between the iris and the trabecular meshwork.

[0054] In another embodiment of the present invention, the lens morphology analysis in the 2D result includes at least one of Front R, Thickness, Diameter, Decentration, and Tilt.

[0055] Front R (Front Radius) is the radius of curvature of the front surface of the crystalline lens in the 2D tomographic image (mm).

[0056] Back R (Back Radius) is the posterior radius of curvature of the lens (mm) in the 2D tomographic image.

[0057] Thickness (Thickness) is the thickness (mm) of the crystalline lens in the 2D tomographic image.

[0058] Diameter (Diameter) is the diameter (mm) of the crystalline lens at the equator in the 2D tomography image.

[0059] Decentration (Decentration) is the amount of decentering (mm) of the central axis of the lens from the axis passing through the corneal apex (Vertex Normal) in a 2D tomographic image.

[0060] Tilt (Tilt) is the inclination (°) of the central axis of the crystalline lens with respect to the axis passing through the corneal apex (Vertex Normal) in the 2D tomographic image.

[0061] In another embodiment of the present invention, the lens morphology analysis in the 3D result includes at least one of Front R, Front Rs, Front Rf, Back R, Back Rs, Back Rf, Thickness, Diameter, Decentration, and Tilt.

[0062] Front R (Front Radius) is the average value (mm) of the strongest meridian Rs and weakest meridian Rf of the radius of curvature of the front surface of the crystalline lens in 3D.

[0063] Front Rs (Front Radius Steep) is the steepest meridian Rs (mm) and axial angle (°) of the 3D front radius of the lens.

[0064] Front Rf (Front Radius Flat) is the weakest meridian Rf (mm) and axial angle (°) of the 3D front radius of the lens.

[0065] Back R (Back Radius) is the average value (mm) of the strongest meridian Rs and weakest meridian Rf of the posterior radius of curvature of the crystalline lens in 3D.

[0066] Back Rs (Back Radius Steep) is the steepest meridian Rs (mm) and axial angle (°) of the posterior radius of curvature of the lens in 3D.

[0067] Back Rf (Back Radius Flat) is the minor meridian Rf (mm) and axial angle (°) of the posterior radius of curvature of the lens in 3D.

[0068] Thickness (Thickness) is the thickness of the lens in 3D (mm).

[0069] Diameter (Diameter) is the 3D diameter (mm) of the crystalline lens at its equator.

[0070] Decentration (Decentration) is the amount of decentering (mm) and the axis angle (°) of the central axis of the lens from the axis passing through the corneal apex (Vertex Normal) in 3D.

[0071] Tilt (Tilt) is the inclination (°) of the central axis of the crystalline lens with respect to the axis passing through the corneal apex in 3D (Vertex Normal).

[0072] In another embodiment of the invention, the lens morphology analysis includes at least one of Ks, Kf, CYL, ACCP, ECC, AA, Apex, Thinnest, and ESI.

[0073] Ks (Keratometry Steep) is the steepest meridian (D or mm) which is synonymous with K2 in a keratometer.

[0074] Kf (Keratometry Flat) is the weak principal meridian (D or mm) which is synonymous with K1 in a keratometer.

[0075] CYL (Cylinder) is corneal astigmatism (D).

[0076] ACCP (Average Central Corneal Power) is the average corneal refractive power (D) within a diameter of 3 mm.

[0077] ECC (Eccentricity) is the corneal eccentricity.

[0078] AA (Analyzed Area) is the percentage of the area available for corneal morphology analysis.

[0079] Apex (Apex) is the central corneal thickness (μm) in the corneal thickness map.

[0080] Thinnest (Thinnest) is the thickness (μm) of the thinnest part in the corneal thickness map.

[0081] The Ectasia Screening Index (ESI) is a screening index for keratoconus.

[0082] In another embodiment of the present invention, the reference points include at least one of SS, AR, IR, IRT and EP.

[0083] SS (scleral spur) is the scleral promontory, AR (angle recess) is the bottom of the angle, IR (iris root) is the root of the iris, IRT (iris rear tip) is the rear tip of the iris (contact end point of the lens), and EP is the end point of the angle (ITC end point).

[0084] In another embodiment of the invention, the corneal configuration data set includes at least one of DSI, OSI, CSI, SD_P(4mm), CV_P(4mm), ACP(3mm), RMS_E(4mm), SR_E(4mm), SR_H(4mm), CSI_T, SD_T(4mm) and CV_T(4mm).

[0085] As shown in Figure 8, the φ9 area of ​​the cornea is divided into eight equal sectors. DSI (Differential Sector Index) is defined as the difference calculated by subtracting Min_Area_Power [D] from Max_Area_Power [D]. OSI (Opposite Sector Index) is defined as the difference calculated by subtracting Opposite_Area_Power [D] from Max_Area_Power [D]. Max_Area_Power [D] is defined as the maximum refractive power (Axial Power). Min_Area_Power [D] is defined as the minimum refractive power (Axial Power). TIFF0007660805000001.tif1191 TIFF0007660805000002.tif9100 Opposite_Area_Power[D] is defined as the axial power in a sector located opposite to the sector having the maximum axial power.

[0086] CSI (Center Surround Index) is defined as the difference calculated by subtracting OuterP[D] from InnerP[D], where InnerP[D] is the refractive power (axial power) in the central region (φ0 to 3 mm) and OuterP[D] is the refractive power (axial power) in the peripheral region (φ3 to 6 mm), as shown in Figure 9. TIFF0007660805000003.tif847

[0087] SD_P(4mm) (Standard Deviation of Corneal Power (φ4)) is defined as the standard deviation of the axial power data within the φ4 region of the cornea, calculated by the following formula: TIFF0007660805000004.tif2267 N is the number of axial power data, P is the axial power data, TIFF0007660805000005.tif30150 is the average value of axial power.

[0088] CV_P(4mm) (Coefficient of Variation of Corneal Power(φ4)) is defined as the coefficient of variation of the axial power data within the φ4 region of the cornea, calculated by the following formula: TIFF0007660805000006.tif1570

[0089] ACP(3mm) (Average Corneal Power(φ3)) is defined as the average value of the axial power data within the φ4 region of the cornea, calculated by the following formula: P(i) is the measurement point, and PatchArea(i) is the area of ​​the measurement point (Figure 10). TIFF0007660805000007.tif2482

[0090] RMS_E(4mm) (Root Mean Square of corneal Elevation (φ4)) is defined as the root mean square of the elevation data within the φ4 region of the cornea, calculated by the following formula: TIFF0007660805000008.tif2154N is the number of elevation data and E is the elevation data.

[0091] SR_E(4mm) (Surface Regularity of Corneal Elevation (φ4)) is defined as the surface regularity index (SRI) based on elevation data within the φ4 region of the cornea, calculated by the following formula (Figure 11). TIFF0007660805000009.tif17150 TIFF0007660805000010.tif20150 TIFF0007660805000011.tif14150 TIFF0007660805000012.tif8150

[0092] SR_H(4mm) (Surface Regularity of Corneal Height (φ4)) is defined as the surface regularity index based on the height data within the φ4 region of the cornea, calculated by the following formula (Figure 12). TIFF0007660805000013.tif2095 TIFF0007660805000014.tif3395

[0093] The Center Surround Index of Thickness (CSI_T) is defined as the difference between the average thickness of the central region (φ3) and the average thickness of the peripheral region (φ6) (FIG. 13), calculated by the following formula: TIFF0007660805000015.tif965

[0094] SD_T(4mm) (Standard Deviation of Thickness(φ4)) is defined as the standard deviation of corneal thickness (pachymetry) within the φ4 region of the cornea, calculated by the following formula: TIFF0007660805000016.tif2164

[0095] CV_T(4mm) (Coefficient of Variation of Thickness(φ4)) is defined as the coefficient of variation of corneal thickness (pachymetry) within the φ4 region of the cornea, calculated by the following formula: TIFF0007660805000017.tif1468

[0096] In another embodiment of the invention, the tomography is optical coherence tomography.

[0097] In another embodiment of the invention, the optical coherence tomography is anterior segment optical coherence tomography.

[0098] In another embodiment of the invention, the optical coherence tomography is swept-source optical coherence tomography. Effect of the Invention

[0099] The proposed methodology resulted in four clusters; I: a cluster consisting of mostly normal eyes (224 eyes with ESI equal to zero, 23 eyes with ESI between 5 and 29, and 9 eyes with ESI greater than 29); II: a cluster consisting of mostly healthy eyes and eyes with incomplete keratoconus (1772 eyes with ESI equal to zero, 698 eyes with ESI between 5 and 29, and 117 eyes with ESI greater than 29); III: a cluster consisting of mostly eyes with mild keratoconus stages (184 eyes with ESI greater than 29, 74 eyes with ESI between 5 and 29, and 6 eyes with ESI equal to zero); IV: A cluster consisting of mostly advanced keratoconus stages (80 eyes had ESI above 29 and 1 eye had ESI between 5 and 29) was identified. According to the present invention, the condition and severity of keratoconus can be successfully identified using unsupervised machine learning algorithms along with linear and non-linear corneal data transformations. The proposed method allows for better identification and visualization of corneal disease stages. [Brief description of the drawings]

[0100] [Figure 1] Principal component analysis applied to corneal features: left, variance revealed for the first 40 significant principal components; right, corneal features in the space of the first 6 principal components. [Diagram 2]Evolution of corneal parameters in 2-D tSNE space. From left to right, we zigzag the evolution of tSNE over time starting from an initial state, where the corneal parameters are simply scattered in the 2-D space, and then grouping eyes with similar corneal properties over time. [Diagram 3] Unsupervised machine learning identified eyes with similar corneal characteristics into four clusters. [Figure 4] FIG. 11 is a diagram of the mapping of ESI indices to clustering: on the left, ESI indices corresponding to the anterior cornea, in the center, ESI indices corresponding to the posterior cornea, and on the right, the overall ESI index of the Casia instrument. [Diagram 5] Graphs for an investigation of another density-based clustering algorithm, CLUTO: top left, applying CLUTO to the tSNE-specific parameters and visualized on a tSNE map; top right, applying CLUTO to the PCA components and visualized on a tSNE map; bottom left, applying CLUTO to the original data with 420 parameters and visualized on a tSNE map; bottom right, applying CLUTO to the original data and visualized using the two significant principal components. [Figure 6] FIG. 1 is a diagram illustrating the concept of 2D analysis. [Figure 7] FIG. 1 is a diagram illustrating the concept of STAR360° 2D analysis. [Figure 8] FIG. 1 shows the φ9 region of the cornea divided into eight equal sectors. [Figure 9] FIG. 1 is a diagram illustrating the concept of CSI. [Figure 10] FIG. 1 is a diagram illustrating the concept of measurement points. [Figure 11] FIG. 1 is a diagram illustrating the concept of elevation. [Figure 12] FIG. 1 is a diagram illustrating the concept of height. [Figure 13] FIG. 1 is a diagram illustrating the concept of CSI_T. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0101] In this multifaceted and retrospective study, corneal optical coherence tomography (OCT) images from 12,242 eyes of 3162 subjects using the SS-1000 CASIA OCT imaging system (Tomey Corporation, Japan) and other parameters from the electronic health record (EHR) system were collected. All data available on each device were collected without any preconditioning. Then, a single visit from each eye was selected, and eyes without an Ectasia Status Index (ESI) were excluded. A total of 3,156 eyes met the criteria. Approximately 57% of participants were female, and the mean age was 69.7 years (standard deviation; SD = 16.2). Three screening labels were obtained from the ESI index of Casia (Tomey Corporation). If the ESI is between 0 and 4, it is normal, if the ESI is between 5 and 29, it is incomplete keratoconus (or suspected keratoconus), and if the ESI is 30 or more, it is keratoconus. Using the Casia label, this dataset included 1970 healthy eyes, 796 eyes with incomplete keratoconus, and 390 eyes with keratoconus. The ESI is essentially an instrument-guided screening index, and has been shown to have good agreement with the Belin-Ambrosio (BA) index in the diagnosis of keratoconus (Non-Patent Document 12). This study was conducted and submitted in accordance with the Declaration of Helsinki and the ethical standards of the Institutional Review Board (IRB), and was approved by the IRB Office of Jichi Medical University. A data use agreement to conduct the analysis was signed between the Japanese center and this institute. The data was anonymized in Japan before further processing.

[0102] For the unsupervised machine learning analysis, 420 parameters were selected, including the axis, refraction, elevation, and pachymetry of both the anterior and posterior corneal surfaces. All ESI-related parameters were excluded from the dataset. First, a principal component analysis (PCA) was applied to the 420 selected corneal parameters using the prcomp function in the R package. PCA uses linear and orthogonal transformations to convert the observed values ​​of highly correlated corneal parameters into a new set of parameters that are not linearly correlated with each other. In other words, each new principal component parameter is a weighted combination of all the initial corneal parameters, while the components are no longer correlated. This transformation allowed a linear reduction in the dimensionality of the original dataset. To investigate how many principal components are significant compared to the null distribution we generated, we generated 100 independent artificial datasets and randomly permuted the values ​​along each corneal parameter within each dataset (Non-Patent Document 13). This operation removes pairwise correlations between corneal parameters while keeping the distribution of each parameter unchanged. We then applied PCA to each of these 100 synthetic VF datasets to classify the combined eigenvalues ​​of the different datasets. We identified principal components in datasets whose eigenvalues ​​were significantly larger (p<0.01, Bonferroni correction) than the top eigenvalues ​​of the synthetic datasets.

[0103] Next, we applied diversity learning using t-distributed stochastic neighbor embedding (tSNE) (see Non-Patent Document 14) to group eyes with similar corneal characteristics together and separate eyes with different characteristics as far as possible. For this purpose, we used the Rtsne function in the R package. This method maps eyes with similar local distance measures in the tSNE space, non-linearly reducing the dimensionality of the input data. Furthermore, tSNE provides a user-friendly visualization, making it suitable for visualizing and monitoring the progression of the disease by clinicians. Furthermore, tSNE allows subjective validation of the subsequent unsupervised clustering, since we can see how the clusters are distributed and overlapping in two- or three-dimensional space. More importantly, tSNE compares the best two principal components to generate more distinct, non-overlapping clusters.

[0104] While there are several unsupervised clustering algorithms to identify hidden structures in a dataset (Non-Patent Documents 15-20), we used unsupervised density-based clustering in tSNE space (see Non-Patent Document 21) to identify eyes with similar corneal characteristics in tSNE space and objectively group these eyes into non-overlapping clusters. Density-based clustering groups eyes that are closely clustered together in tSNE space and have many neighbors around them, while eyes that are located alone (in low-density regions) and quite far away are marked as outliers and not part of the group. We then assessed the accuracy of the method qualitatively (by visualization) and quantitatively (using the screening index of the Casia instrument).

[0105] Figure 1 (left) shows the top 40 principal components and the amount of variance in the data that these components revealed. Based on quantitative analysis, 32 principal components were identified as significant. The top 32 principal components revealed more than 80% of the total data variance in the data, while the top 8 principal components retained approximately 60% of the total data variance in the data. However, after examining the tSNE map, we selected 8 principal components that showed more distinctive clusters on the tSNE map (qualitative validation). Two corneal-specific parameters were generated that are essentially nonlinear combinations of the original corneal parameters.

[0106] Figure 2 shows the evolution of tSNE over time starting from an initial state where the corneal parameters are scattered in 2-D space without considering local properties among the points. 2-D was chosen because it provides a clear visualization for the user, and importantly the clinician. In this case, the algorithm identifies eyes with similar properties based on their distance in the tSNE space and gradually groups these eyes together. The perplexity, reflecting the expected number of neighbors for each point, was set to 34, allowing a maximum number of iterations up to 1000.

[0107] To subjectively evaluate the learning accuracy, unsupervised density-based clustering was applied to the two identified corneal intrinsic parameters. Clusters with less than seven eyes were excluded. Unsupervised density-based clustering identified four non-overlapping clusters. For better visualization, the clusters were color-coded as shown in Figure 3.

[0108] We then assigned clinical labels to the four identified clusters based on the ESI index (ranging from 0 to 100) provided by the Casia instrument, where zero indicates normal and 100 reflects the most advanced stage of keratoconus. The Casia instrument also provides diagnostic labels based on the ESI index: normal when ESI equals zero, incomplete keratoconus (or suspected keratoconus) when ESI is between 5 and 29, and keratoconus when ESI is greater than 29. However, it is not clear how this index was generated from all corneal parameters, and more importantly, how the threshold for identifying eyes with incomplete keratoconus was identified. Furthermore, the currently used incomplete keratoconus threshold index is inherently confusing, since keratoconus represents a wide range of corneal deformations, especially in the early stages, and it is difficult to assign a binary label that separates normal eyes from eyes with incomplete keratoconus. Nevertheless, using Casia's ESI index and diagnostic label conventions, it was determined that cluster I (coded in black) consisted mainly of healthy eyes: 224 healthy eyes, 23 eyes with incomplete keratoconus, and 9 eyes with keratoconus. Cluster II (coded in grey - the large cluster on the left) consisted mainly of healthy eyes and eyes with incomplete keratoconus: 1772 healthy eyes, 698 eyes with incomplete keratoconus, and 117 eyes with keratoconus. Cluster III (coded in light grey) consisted mainly of eyes with mild keratoconus: 184 eyes with mild keratoconus, 74 eyes with incomplete keratoconus, and 6 healthy eyes. Small cluster IV (coded in black with white triangles) consisted mainly of eyes with advanced keratoconus: 80 eyes with advanced keratoconus and 1 eye with incomplete keratoconus.

[0109] To subjectively assess the correlation between the severity of keratoconus and the ESI index of the Casia device for eyes in the identified clusters, each eye was color-coded on the clustering graph by the anterior ESI index, posterior ESI index, and total ESI index, which reflect the severity of keratoconus. Figure 4 shows the mapping of the anterior ESI index, posterior ESI index, and total ESI index to the identified clusters.

[0110] To objectively evaluate the accuracy of unsupervised clustering, we calculated the specificity and sensitivity based on the Casia diagnostic labels. The specificity for identifying healthy eyes from eyes with keratoconus was 94.1%, and the sensitivity for identifying eyes with keratoconus from healthy eyes was 97.7%. To compare the DBSCAN clustering algorithm against other methods, we investigated the OPTICS (Non-Patent Document 19) and Clustering Toolkit (CLUTO) algorithms (Non-Patent Document 20). CLUTO is a software package for unsupervised clustering on low- and high-dimensional datasets. First, we applied CLUTO on tSNE and visualized the results. We then sought to determine whether CLUTO would have produced more discriminatory clusters using the principal components, or the original data with 420 parameters. Figure 5 shows how CLUTO clustered the eyes using tSNE-specific parameters, principal components, and the original data. As can be seen, none of the results produced well-separated clusters. To objectively evaluate the clustering results, we further investigated the specificity and sensitivity of the clusters using the same methodology implemented for DBSCAN. We found that DNCLUE produced four clusters that were typically normal, and one cluster that was abnormal. We implemented OPTICS and CLUTO using the optics and skmeans functions in R, respectively. To objectively examine the clusters generated by the CLUTO algorithm, we calculated the specificity and sensitivity of CLUTO applied to the original data with 420 parameters. The specificity for identifying healthy eyes from those with keratoconus was 97.4%, and the sensitivity for identifying eyes with keratoconus from healthy eyes was 96.3%. However, DBSCAN applied on tSNE was chosen because this combination resulted in acceptable accuracy and well-separated clusters matching different stages of keratoconus.

[0111] The main finding of the present study is that an automated unsupervised clustering algorithm using corneal topography, tomography and thickness profiles provides a specific and sensitive means for determining the status and severity of keratoconus. The proposed unsupervised machine learning analysis for keratoconus diagnosis and staging provides a promising tool to improve early detection of keratoconus at its early stages and enable monitoring of disease treatment.

[0112] Marc Amsler first described how keratoconus manifests as an altered corneal topography in 1938. However, the introduction of computer-assisted videokeratoscopy in the early 1980s revolutionized the diagnosis of keratoconus. Since then, most of the early methods for identifying keratoconus and severity indexes have been based on corneal topography (see non-patent literature 2, 4, 9, and 22-25). More recently, based on a cohort of 44 eyes with keratoconus and 113 healthy eyes, we have shown that the pachymetry index can successfully distinguish healthy eyes from those with keratoconus (non-patent literature 26). However, in the present study, we used corneal topography, elevation, and thickness profiles extracted from optical coherence tomography (OCT) images of the subjects using the SS-1000 Casia to identify and stage keratoconus.

[0113] Historically, classification of keratoconus stages has been based on qualitative analysis of the morphology of the entire cornea. However, in the present invention, machine learning was used because it addresses limitations to currently used diagnostic methods, including the assessment of parameters that are qualitative rather than quantitative, and observer bias. Machine learning algorithms for keratoconus have been proposed previously, but most are based on a single type of corneal parameter (e.g., topography alone) (Non-Patent Documents 23-25) or require pre-labeled data (Non-Patent Documents 4, 7, and 27). For example, some researchers have used supervised neural networks or tree-based classification to distinguish between normal eyes and eyes with keratoconus (Non-Patent Documents 4, 27-29). However, subjectively pre-labeling an eye as keratoconus or incomplete keratoconus is itself subject to subjective assessment and bias.

[0114] We used approximately 420 corneal parameters generated by the Casia instrument through swept-source OCT images of the cornea. All of these corneal parameters were transformed into 2-D space using linear PCA and nonlinear tSNE, followed by an unsupervised machine learning algorithm. Thus, instead of feeding all the parameters to the machine learning algorithm and perturbing its predictions, we first extract highly predictable corneal condition information. However, most of the machine learning algorithms in the literature only input different corneal parameters into the machine learning algorithm for keratoconus identification, but do not enhance the data transformation power and extract the most useful knowledge for disease identification. To investigate whether PCA alone could produce clusters that were as well separated as those identified by the combination of PCA and tSNE, we applied PCA alone and performed clustering. We found that PCA alone produced clusters with significant overlap. We also applied CLUTO to selected principal components and compared the results with tSNE, observing similar overlapping clusters (Figure 5, top right).

[0115] A subjective assessment of the quality of learning using the cluster visualization and overlaid with the Casia ESI Keratoconus Index (shown in Figure 4) found that the anterior corneal ESI index was highly correlated with the keratoconus severity identified within the cluster. In particular, eyes classified by machine learning as having advanced keratoconus in cluster IV (color coded with white triangles and black) have high agreement in the anterior ESI index, posterior ESI index, and total ESI index, as almost all eyes in this cluster have black.

[0116] The same analogy holds for eyes in cluster I (colored black) that were classified as normal based on machine learning. However, for cluster III (small cluster, colored light gray), which represents mild keratoconus based on machine learning, eyes generally have worse posterior ESI indices compared to anterior ESI indices (Figure 4, left and middle panels). This finding may suggest that posterior corneal parameters better identify keratoconus. However, this finding requires further investigation. We hypothesize that eyes within this cluster that were labeled as normal by the Casia ESI labeling system may be candidates for "incomplete" keratoconus that requires more attention from clinicians. Finally, cluster II (colored dark gray), which represents healthy and suspected keratoconus based on machine learning, has strong agreement with all ESI indices, except for the far right tail. Eyes in this region do not have good agreement with the anterior and total ESI indices. It is hypothesized that this region may be either a separate cluster or part of cluster III and cannot be identified based on the current data and the algorithm used. It is also possible that eyes within this subregion have mild keratoconus along with other ocular conditions that we did not have enough information to characterize. However, Figure 4 (middle panel) shows that the posterior ESI index was more effective than the anterior ESI index (Figure 4, left panel). In fact, a study conducted by the Ambrosio group has shown that posterior features are superior to anterior features in identifying keratoconus (Non-Patent Document 30).

[0117] Within these clusters, the number of eyes with large or small ESI was evaluated to objectively evaluate the clinical labels assigned to the clusters according to the severity of the eye. All eyes in cluster IV (advanced keratoconus by machine learning) had an ESI index greater than 38. In this group, 71 eyes (out of 81 eyes) had an ESI greater than 60, indicating an advanced stage of keratoconus. In cluster I (normal by machine learning), only 7 eyes (out of 256 eyes) had an ESI index greater than 30, indicating that the vast majority of eyes in this cluster were normal. Thus, the clustering according to the present invention is in good agreement with the ESI in the two-sided range. Based on the Casia ESI diagnostic label, the specificity of the machine learning method of the present invention in identifying normal eyes from keratoconus was 94.1%, and the sensitivity of identifying keratoconus from normal eyes was 97.7%, considering only the normal cluster and the advanced keratoconus cluster.

[0118] There are some limitations in the study on the present invention, which can be addressed in subsequent studies. The clustering results were compared with the Casia ESI index, showing good agreement between our findings and the ESI index range (Figures 3 and 4). However, to evaluate the generality of this unsupervised clustering method, it needs to be validated with other keratoconus indices, such as the Bellin-Ambrosio (BA) index. Therefore, another study needs to be performed to see how this method can be generalized to corneal parameters generated by the Pentacam instrument by having access to such a dataset. Also, the accuracy of this method can be verified if clinical diagnostic labels for all eyes are available. However, accessing clinical diagnostic labels for all eyes in such a big dataset is a difficult and tedious task. Nevertheless, it would be beneficial to evaluate the proposed method with a dataset containing clinical diagnostic labels in subsequent studies.

[0119] Qualitative and quantitative evaluations were performed to determine whether PCA alone or other clustering techniques produced sufficiently separated clusters. We found that the OPTICS density-based clustering technique was able to separate eyes with different keratoconus stages, while the CLUTO unsupervised clustering technique produced overlapping clusters. However, the most important aspect of our proposed technique is the visualization properties and tSNE 2-D maps. This is important in a real clinical setting, making it more suitable for monitoring disease progression on 2-D maps than the black box we propose without 2-D visualization. In summary, the present invention has proposed a possible solution to address the shortcomings of current keratoconus diagnosis and monitoring approaches, including observer bias in pre-defined diagnoses and the limitation of only providing binary results, such as whether the eye belongs to either a normal or diseased group. The unsupervised machine learning algorithm introduced does not require pre-labeled data for training and can automatically identify the keratoconus status for a given eye based on comprehensive corneal parameters including topography, elevation and thickness profiles. More importantly, this unsupervised machine learning algorithm provides visualization of the eye's status compared to other eyes in different keratoconus stages, which was not present in supervised machine learning methods. To the best of the inventor's knowledge, this is the first attempt to deploy a fully unsupervised algorithm for keratoconus identification and monitoring.

[0120] Indeed, the condition and severity of keratoconus can be successfully identified using an automated, unsupervised clustering algorithm that uses corneal topography, tomography and thickness profiles. This approach can be used in corneal clinical and research settings to better diagnose, monitor changes and progression, and improve understanding of corneal changes in keratoconus.

[0121] It should be appreciated that the present invention can be used in many different ways and for many different applications to dramatically improve the identification of corneal severity.

Claims

1. obtaining a corneal configuration data set of the tomographically examined cornea; and Visualizing the corneal configuration data set of the examined cornea in a two-dimensional or three-dimensional map using principal component analysis and t-distribution type probabilistic neighborhood embedding in that order together with several existing corneal configuration data sets of diseased, suspected and normal corneas obtained by the tomography. A method for assisting in identifying the severity of the examined cornea, comprising:

2. 2. The method of claim 1, wherein the severity is one selected from the group consisting of keratoconus severity, bullous keratopathy severity, corneal haze severity, corneal leukoplakia severity, corneal ulcer severity, corneal herpes severity, corneal chemical burn severity, corneal thermal injury severity, and corneal degeneration severity.

3. The method of claim 1 or 2, wherein the tomography is optical coherence tomography.

4. The method of claim 1 or 2, wherein the tomography is anterior segment optical coherence tomography.

5. The method of claim 1 or 2, wherein the tomography is swept-source optical coherence tomography.

Images